Although this may be the year that a supercomputer achieves the one petaflop Linpack performance mark, I mostly see 2008 as a consolidation period for high performance computing. The important technologies that emerged in 2007 — quad-core processors, DDR InfiniBand, coprocessor accelerator technology and a variety of new parallelizing software tools (from Intel, Microsoft, RapidMind, Interactive Supercomputing and others) — will become more fully utilized in 2008. Over the next year, system makers and users will start to figure out how to take advantage of the additional cores, accelerators, faster interconnects and new software technologies to build more productive systems.
These existing technologies will be used not just to create a few petaflop supers, but also a whole lot of sub-$50,000 systems for department and workgroup teams. According to IDC and others, this latter segment represents the biggest opportunity for HPC vendors for 2008 and beyond. In this issue, John West’s feature article takes a closer look at this much talked-about HPC mid-market opportunity and how the vendors are approaching it.
If the high volumes are truly at the low end, that would suggest that we’re heading for the era of the “personal supercomputer” — a much maligned term denoting a sub-$10,000 HPC workstation. As long as the utility computing model doesn’t swallow up supercomputing in the next few years, HPC workstations could indeed become the next sweet spot in HPC. Probably not for 2008 though. Even though a lot of the pieces already exist — powerful multicore CPUs and GPUs, standard software libraries, high performance interconnects and a choice of open source or commercial clustering middleware — no one has been able to glue it all together yet into a compelling product. (SiCortex may have the right idea with their Catapult, which is essentially a “mini-me” version of their larger production machines.) Since we’re still two or three years away from manycore processors, integrating one or more accelerators into a workstation is going to be necessary if you’re after teraflop-level performance.
But I’m getting ahead of myself.
Even though 2008 may be a year for consolidation, a raft of exciting products are in the pipeline. Scheduled to debut this year are eight-core x86 chips (in Intel’s Nehalem family), Windows HPC Server 2008, 64-bit GPUs for computing, QDR InfiniBand and optical cables for cluster interconnects, just to name a few. A host of HPC-friendly 10GbE products are also waiting to do battle with InfiniBand. Even assuming all these products arrive on schedule, their real impact probably won’t be realized until 2009, after the early adopters have kicked the tires for a few months.
On the high performance processor front, the big story will continue to be the battle between Intel and AMD. With Intel completely dominating its smaller rival in 2007, the company is looking to land a death blow with its new Nehalem processor family in the second half of 2008. Nehalem is the “tock” microarchitecture redesign of the 45nm Penryn “tick.” As a scalable architecture containing 2 to 8 processor cores, one or more of which could be a GPU, Nehalem might get a jump on AMD’s CPU-GPU “Fusion” processors (now called Accelerated Processing Units). Since AMD has pushed its first CPU-GPU offerings into the second half of 2009, Intel could conceivably have a Nehalem CPU-GPU alternative shipping by then.
First though, Intel will probably need to figure out how to get four cores (at least) on the same piece of silicon. Their current strategy of packaging two dual-core chips for their quad-core products will likely not extend to an eight-core Nehalem socket. When you include the additional complexity of an integrated memory controller and Intel’s new QuickPath interconnect, the Nehalem architecture represents the biggest change to the company’s processor design since they launched the 8086. But if Intel executes its Nehalem roll-out flawlessly (a big if), it will make it almost impossible for AMD to regain the x86 high ground for the foreseeable future.
Intel’s high-end GPU play is “Larrabee,” a general-purpose graphical computing processor architecture, scheduled to be demo’ed later this year (but probably not delivered until 2009). Meanwhile, NVIDIA, undeterred by having to worry about those pesky CPUs, is scheduled to introduce its next-generation Tesla products for GPU computing early in the new year. Those devices will presumably support 64-bit floating point math, bringing them closer to a general-purpose commodity vector processor for HPC. AMD has its own 64-bit graphics computing offering in the FireStream stream processor, also scheduled for delivery in Q1 2008.
With all the GPU talk in 2007, FPGAs have faded somewhat from the limelight. But they still represent perhaps the biggest unrealized potential of any HPC processor technology. As such, FPGAs seem forever poised to achieve greatness. Over the last year, it’s become much easier to connect these chips into standard servers, thanks to AMD and Intel providing the necessary hooks into the host processor. Software development remains the biggest challenge to more widespread adoption. If one of the two leading FPGA vendors, Xilinx or Altera, were to get directly involved in developing HPC-specific FPGA products themselves, like NVIDIA did for GPUs, I might be more bullish on these devices going mainstream. Until then, I see FPGAs mainly exploiting niches where the advantages of performance plus reconfigurability outweigh the extra programming pain.
Network interconnects will be another area to watch in 2008 as users look for a unifying fabric to connect cluster and storage nodes. In 2007, the popularity of InfiniBand continued to expand, riding the growth of high performance clusters. In 2008, iWARP NICs and faster switches may finally give 10GbE the boost it needs to challenge InfiniBand. But I’m still skeptical. The Ethernet crowd thinks its time has come, but I think they’ve underestimated the maturity and momentum built up by InfiniBand over the last few years. On the other hand, it’s tough betting against Ethernet in the long term. 2008 may provide some hints of how this will play out. (I delved a bit deeper into this topic in the final issue of 2007; see InfiniBand and 10GbE Head for Showdown).
If that’s not enough prognosticating for you, check out our two feature articles from the HPC brain trusts at IDC and Tabor Research, which offer some of their HPC predictions for the upcoming year. Both analyst teams run the numbers, estimating the size of the total HPC market for 2008. IDC continues to see the largest demand for medium to small HPC systems for departments and workgroups, while Tabor Research thinks a new category of HPC “threshold applications” will provide major growth. Both groups end up with a similar dollar figure for the total HPC market — in the neighborhood of $20 billion.
OK, $20 billion sounds like a lot. But what if a recession hits in the U.S.? The general consensus from economists seems to be that there’s about a 50/50 chance for one in 2008. If a credit crunch drives the recession, as many suspect it will, IT investments will nosedive. The impact on these high-growth HPC application areas could be substantial. The fallout might eliminate some struggling vendors as well as make others more susceptible to acquisition and mergers. That’s not necessarily the Doomsday scenario it suggests. A leaner HPC vendor community might serve to focus talent and resources at companies where they could do the most good.
Recession or not, HPC seems destined to continue its penetration into technology-hungry business segments over the next year. At the same time, high-end supercomputers will almost certainly establish new performance records and push the application envelope. All in all, 2008 should provide plenty of interest and intrigue for high performance computing users and spectators alike.
As always, comments about HPCwire are welcomed and encouraged. Write to me, Michael Feldman, at [email protected].